Sei sulla pagina 1di 29

Economics 314 Coursebook, 2010 Jeffrey Parker

5 THEORIES OF ENDOGENOUS GROWTH

Chapter 5 Contents

A. Topics and Tools ............................................................................. 1 


B.  The Microeconomics of Innovation and Human Capital Investment ........... 3 
Returns to research and development ..............................................................................4 
Human capital vs. knowledge capital .............................................................................6 
Returns to education .....................................................................................................6 
C. Understanding Romer=s Chapter 3, Part A ............................................. 9 
Introduction .................................................................................................................9 
Romer’s modeling strategy and the research literature ......................................................9 
The basic setup of the R&D model ............................................................................... 10 
The knowledge production function .............................................................................. 11 
Analysis of the model without capital ........................................................................... 12 
The R&D model with capital ...................................................................................... 16 
Returns to scale and endogenous growth ....................................................................... 17 
Scale effects in the R&D model .................................................................................... 18 
D.  Understanding Romer’s Chapter 3, Part B ........................................... 20 
The specification of the human-capital model ................................................................ 20 
Analysis with the human-capital model ........................................................................ 20 
Output per person vs. output per worker........................................................................ 21 
The specification of Romer’s predation model ................................................................ 24 
Agents’ decisions in the predation model ....................................................................... 25 
E.  Suggestions for Further Reading ........................................................ 28 
General texts on modern growth theory......................................................................... 28 
Selected seminal papers in modern growth theory .......................................................... 28 
F. Works Cited in Text ........................................................................ 29 

A. Topics and Tools


The neoclassical growth theory that we studied in Chapters 3 and 4 largely
evolved in the 1950s. There was considerable filling-in of details in the 1960s, but by
the 1970s growth theory had largely become moribund. A tremendous revitalization
has occurred since the 1980s, spurred by several shortcomings of the previous theo-
ries.
First, because growth rates are exogenous in the Solow and Ramsey models,
these theories are unable to explain why growth rates (and, in particular, the rate of
technological progress) might change from one time period to another. This became
an important research topic in the 1980s when emerging data began to convince ma-
croeconomists that productivity growth in the United States and other advanced
countries had declined significantly beginning about 1974.
A second failing of neoclassical growth theory is that it cannot explain the large
and lasting differentials in per-capita income that we observe across countries and
regions. Solow’s growth model implies more rapid convergence of incomes than
seems actually to have occurred, particularly between developed and developing
countries. International differences in technological capability can help explain this
gap, but beg for an economic explanation that cannot be provided by models in
which technology is exogenous.
Another feature of neoclassical growth models that some economists and poli-
cymakers find troublesome is that they provide no mechanism by which the saving
and investment rate (or government policies directed at influencing it) can affect the
steady-state growth rate. While this conclusion of neoclassical models is not obviously
counterfactual, many find it counterintuitive and have explored models in which sav-
ing plays a more central role.
The pioneer of “endogenous growth theory” is Paul Romer, a former colleague
1
but not a relative of our textbook author. His 1986 paper in the Journal of Political
Economy is a seminal work in the modern revitalization of growth theory. The prin-
cipal engine behind endogenous growth is the elimination of the assumption of de-
2
creasing returns to “capital.” In order to justify this radical departure from a long-

1
In the early 1990s, there were three famous young Romers teaching macroeconomics at the
University of California at Berkeley. Paul, who focuses on growth theory and is now at the
Stanford Business School, David (our author), who is a prominent neo-Keynesian, and Da-
vid’s wife Christina, who is a macroeconomic historian and now chair of the Council of Eco-
nomic Advisors in the Obama Administration.
2
This is a good time to clarify two closely related concepts: “diminishing marginal returns”
and “decreasing returns to scale.” The former is usually applied to changes in only a single
factor of production holding all other factors constant. Thus, diminishing returns to capital
means that when more capital is added to production with all other factors held constant, the
ensuing increase in output becomes smaller as more and more capital is added. Returns to
scale usually apply to the effect on output of simultaneous changes in many or all factors of
production. “Constant returns to scale” by itself means that increases of an equal percentage
in all factors leads to an increase of the same percentage in output. In this chapter, we will
extend the idea of returns to scale to situations where a subset of factors changes.

5–2
established assumption of microeconomic theory, Romer and his followers have
broadened the definition of capital to include human capital and/or knowledge capi-
tal. As we shall see, once this broader view of capital is adopted it is no longer ob-
vious that there are decreasing returns. This leads to radical changes in the conclu-
sions that we derive from models that are otherwise similar to those of Solow and
Ramsey.
There are three basic models developed in the chapter: the R&D model of 3.2
and 3.3, the human-capital model of 3.8, and the production-protection-predation
model of section 3.11. We will give attention to all three.
The mathematical tools used here are largely familiar ones. To keep the analysis
simple, Romer reverts to the simple Solow assumptions about saving (and other stat-
ic resource allocation decisions). The original literature on these models bases deci-
sions on utility and profit maximization, which is more satisfactory, but the dynamic
properties of the model are similar with constant growth rates, so Chapter 3 will
teach you the essential features of the model without all the complicated mathemat-
3
ics that we saw in Romer’s Chapter 2.
As in previous chapters, we will be searching for steady-state balanced growth
paths. To find these, we will usually look for situations in which the growth rates of
the key state variables are constant. In most of the models of this chapter, there will
be two state variables, either physical capital and knowledge capital or physical capi-
tal and human capital. We will use a two-dimensional phase plane that looks on the
surface like the one in the Ramsey model, but is fundamentally different because in
these models both variables are state variables that cannot jump, whereas in the
Ramsey model c was a control variable that could jump vertically to adjust to
changes in economic conditions.

B. The Microeconomics of Innovation and Human


Capital Investment
Romer’s Chapter 3 examines the macroeconomic implications of investment in re-
search and development (innovation) and human capital. However, some of the
most important theoretical issues in modeling these concepts are microeconomic in
nature. The seminal papers in the modern growth literature vary a lot in how careful-

3
Barro and Sala-i-Martin (2004) is a more advanced textbook that looks at more sophisticated
versions of these models. Acemoglu (2009) is a more recent, and more mathematical, treat-
ment. For those interested in learning about them, Econ 454 develops the more complete
models.

5–3
ly they model these microeconomic issues, but Romer’s simplified presentation of
the models largely ignores the microeconomics. In this section, we briefly consider
some of the basic microeconomic issues involved. (Romer discusses some of these
topics in Section 3.4.)
The models of Chapter 3 attempt to make endogenous the “production” of tech-
nology. In the R&D model, an R&D sector produces additions to society’s stock of
technical knowledge. In the second, individuals add to their human capital by spend-
ing time in education rather than producing output.
A key microeconomic issue that underlies this analysis is the question of what
incentive people have to make investments in knowledge or in human capital. Unless
people get utility directly from the process of research or education (which cannot be
ruled out—consider the case of the “professional student”), they will only undertake
these investments if they are able to profit from them sufficiently to justify the oppor-
tunity cost. The opportunity cost of investing in research or education may include
both forgone consumption and the lost alternative opportunity of investing in physi-
cal capital. Thus, if rational agents invest in research or education, then the earnings
from these activities must have an expected present value at least as high as the current
consumption that must be forgone in order to undertake them and as high as the ex-
pected present value of the returns to physical capital investment, since that is also an
alternative use of funds.

Returns to research and development


As Romer discusses on page 116, pure knowledge is nonrival, meaning that the
use of knowledge by one person does not reduce the ability of others to use it. Most
“private” goods in the economy are, by contrast, rival. To clarify the distinction,
4
think about chocolate-chip cookies. Everyone can use the same (non-rival) recipe for
chocolate-chip cookies but everyone cannot use the same (rival) chocolate chips.
As you learned in Econ 201, a competitive market economy (in the absence of
externalities) can lead to the production of the efficient amount of traditional, rival
goods. The market price provides producers and consumers with a scarcity signal
that can lead to efficient resource allocation by equating the marginal social cost of
the good with its marginal social benefit. On the production side, producers equate
price to the marginal production cost. Consumers consume at the level where the
marginal benefit of an additional unit of the good equals the market price.
Nonrival goods such as knowledge can be reused by the same person or shared
with additional people at zero marginal social cost. With marginal cost equal to zero,
efficiency requires that people should consume knowledge at the level where its mar-
ginal social benefit is also zero. But, as with rival goods, utility-maximizing or profit-

4
Some of us will use any excuse to justify thinking about chocolate-chip cookies.

5–4
maximizing users of knowledge will “purchase” it up to the point its marginal benefit
equals the price that is charged. They will voluntarily choose the optimal level of use
(where marginal benefit is zero) only if the price of knowledge is zero. Thus efficien-
cy requires that knowledge, once it has been created, must be distributed freely at a
zero price.
However, if the market price of knowledge is zero, then the market provides no
financial reward for anyone who incurs the research-and-development costs that are
necessary to create it. To provide such incentives, most countries have patent and
copyright laws that grant exclusive (monopoly) intellectual property rights to indi-
viduals who create knowledge. With a patent or copyright, the creator is able to
charge a positive royalty for the license to use knowledge or to earn monopoly profits
on the newly discovered product or process by using it exclusively and prohibiting its
use by others. However, the argument of the preceding paragraph shows that charg-
ing a positive price for the use of a nonrival good such as knowledge leads to ineffi-
ciency, as does the existence of patent-protected monopolies. If individuals must pay
to use knowledge, but the social cost of it using it is zero, then they will choose to use
knowledge at a lower-than-optimal level. Thus, when intellectual property laws work
as intended, they help resolve one problem by encouraging investment in knowledge,
but at the same time they create another by discouraging its use. This argument is
familiar to anyone who has followed the sometimes intense debates about illegal co-
pying of software, works of entertainment such as CDs and DVDs, or about the pric-
ing of pharmaceuticals.
While patents provide substantial protection in some industries, such as chemi-
cals, intellectual-property laws often do not give innovators much protection. For
example, suppose that a company discovers that a particular tool works better if it is
curved than if it is straight. It can attempt to profit from its discovery by patenting the
curved tool. However, there are many ways to curve a tool and it is probably imposs-
ible to gain patent rights on all possible curves that might be beneficially used. Once
the knowledge that curved tools are better becomes public (as it does when a patent
issues), everyone may be able to adopt some variant of the improved technology
without paying a royalty to the inventor. For such nonexcludable kinds of knowledge,
inventors often resort to secrecy in hopes that it will be costly and time-consuming
for competitors to discover or “reverse engineer” the knowledge. When knowledge is
both nonrival and nonexcludable, it qualifies as a pure public good, with all the fa-
miliar resource-allocation problems that public goods entail. Governments often sub-
sidize research and development for branches of knowledge where nonexcludability
makes patent protection ineffective or where wide diffusion of the resulting know-
ledge seems especially important.
The issue of the efficient allocation of resources to research and development is a
central focus of Reed’s Economics 354: The Economics of Science and Technology.

5–5
If you are interested in pursuing additional readings in this area, visit the instructor’s
Web page for a link to a recent reading list.

Human capital vs. knowledge capital


By human capital we mean acquired characteristics that make workers more
productive. Although it encompasses such characteristics as health, strength, and
stamina, the most commonly analyzed sources of human capital are the education,
training, and experience that a worker embodies. Since education and training in-
volve the transmission of knowledge, it might seem like human capital is the same as
the knowledge capital we study in the R&D model.
However, there is a crucial difference. Knowledge capital is potentially a public
good whereas human capital is not. Perhaps the easiest way of distinguishing be-
tween them is to think about the two major roles that most professors play. You see
professors most often in the classroom, where they are imparting existing knowledge
to students. This increases the students’ human capital, but does not create new
knowledge for society. When they are not in the classroom, your professors are likely
to be engaged in research. If successful, this research leads to new knowledge capital
that everyone can potentially share on a nonrival basis. Thus, simply put, society’s
knowledge capital is everything that is known by someone in the society; your hu-
man capital includes your personal familiarity with and ability to use part of that
knowledge. Your human capital is personal to you—the fact that you have obtained
knowledge may make you more productive but it does not usually raise anyone else’s
productivity. Thus human capital does not have the public-good characteristics of
knowledge capital.

Returns to education
Although human capital is not a public good in the same way as knowledge capi-
tal, education raises interesting economic issues of its own. Some aspects of educa-
tion have elements of nonrivalry. The syllabus for a course or a recorded lecture can
be shared widely at minimal cost. However, most other aspects of education are ri-
val. Classroom seats and instructor time are limited and putting one student into a
seat denies that seat to someone else. Moreover, most kinds of education are easily
excludable. Those who do not pay for a seat in the class can be denied access. Thus,
5
it does not appear that education is a public good.
One can imagine an uncomplicated world in which markets could allocate edu-
cation efficiently. If the benefits of a person’s education and training (including on-

5
Some economists argue that everyone gains from having a more educated society, so that
additional education for one individual benefits others as well as himself. In this case, there is
a positive externality and the market system of incentives will lead to underconsumption of
education.

5–6
the-job training) are perfectly reflected in his or her enhanced productivity, then
someone who has acquired more human capital should receive commensurately
higher wages. In this case, the individual can make an optimal personal decision
about whether the returns to further human-capital acquisition (higher productivity
and wages) justify the cost. There is no market failure here and education/human
capital is similar to other kinds of investment/capital.
However, there are several problems that may upset efficient resource allocation
in education markets. One is the problem of borrowing to finance education invest-
ment. Investments in capital—whether physical, human, or “knowledge capital” ac-
quired through research and development—require a substantial initial expenditure,
followed by a lengthy period over which the investment earns a return. The person
desiring to make the investment often does not have sufficient liquid funds at the
time the investment is to be made, so a “capital market” in which one can borrow for
such expenditures is a useful social institution. However, capital markets can only
function if lenders can be reasonable sure that they will be repaid.
A borrower who purchases a physical capital good such as a building or a ma-
chine must normally pledge the capital good as collateral on the loan. If the borrower
fails to repay the loan as required, the lender can seize the capital good and resell it
to recover at least part of his or her money. However, a borrower’s education cannot
be seized and, in societies that outlaw slavery, borrowers themselves cannot be
6
seized by the lender either. This limits the recourse of lenders in cases of default,
which makes it hard for the private market to provide access to loans for human-
capital investment.
Government-subsidized student loans attempt to remedy this market failure by
providing government guarantees in place of collateral. Many Reed students can
confirm that this allows a thriving market in student loans, but it does not assure al-
locative efficiency. Government guarantees generally make student loans less risky to
lenders than the intrinsic economic riskiness of the underlying returns to education.
Thus, interest rates will usually be subsidized below the level that would be appro-
priate to the investment’s risk and this will encourage overuse of student loans. Fur-
thermore, while the government guarantees allow the market to function, the gov-
ernment is often no more effective at collecting money from defaulters than a private
lender. This diverts the cost of human capital from the investor/student onto the
general taxpayer.
A second difficulty in human-capital investment arises when human capital is
acquired through on-the-job learning. In most jobs, the worker learns a great deal
about how the job is done during the initial months of employment. During that pe-
riod, productivity increases rapidly as the worker gets better at what he or she does.

6
Notwithstanding the fellow who was “repossessed” after failing to pay his exorcist’s bill.

5–7
A “perfect” market might capture this learning by starting the individual at an ex-
tremely low wage (or the new worker might even pay the firm for the privilege of
learning the job), then increasing the wage as productivity rises. This scheme impli-
citly or explicitly makes the worker pay for the investment in human capital. To the
extent that workers value the human capital they acquire, they may be willing to in-
cur this cost, although if the initial wage is low enough it might force them into a
borrowing situation that raises the same problem of collateral described above.
However, much of the knowledge acquired on the job may be “firm-specific”
human capital, such as knowledge of the internal rules and operations of a particular
organization, and be largely useless if a worker moves to another firm. In a world in
which layoffs and job changes are common, workers will be reluctant to bear the cost
of training that is useful only when he or she works for one particular employer. The
firm will also hesitate to invest in a particular worker when the worker might depart,
though there is less risk of the worker quitting at the end of the training period if
most of the training is firm-specific. These difficulties in appropriating the returns to
human-capital investment can lead to underinvestment in training.
A third problem that complicates the efficient allocation of resources to human
capital is that the link between education and productivity is not well understood. It
is uncontroversial that more highly educated workers are more productive; what is at
issue is whether people who are innately more productive tend to invest in more
education or whether it is the education itself that makes them productive. In our
growth models, we assume that education makes individuals more productive. How-
ever, some economists argue that education acts mostly as a screening or ranking
device. According to this “signaling” theory, firms hire college graduates at high
wages not because they have learned anything that makes them more productive, but
because the fact that they finished college signals that they are individuals of high
ability and potentially high productivity.
If one takes this signaling argument to its extreme, then one may claim that edu-
cation has little effect on productivity; it just acts like an elaborate placement test for
employers. For example, a century ago a relatively small share of people finished
high school and very few finished four years of college. According to the signaling
model, being a high-school graduate at that time signaled that you were a high-
quality worker and being a college graduate signaled that you were in an elite catego-
ry of high achievers. Today the majority of people finish high school, so the signaling
value of a high-school diploma is small. Many individuals finish college, so even a
college degree is no indicator of exceptional ability. To demonstrate a truly elite sta-
tus one must now attend graduate school and get an advanced degree. According to
the extreme version of the signaling theory, the people who now get good jobs with a
graduate degree used to be able to get the same jobs (and do them just as well) with a

5–8
bachelor’s degree. If the additional years of study do not raise productivity, then they
are a costly waste of resources.
Although most economists believe that education makes individuals more pro-
ductive, it is difficult to disprove the signaling model because in many cases the two
models lead to similar predicted outcomes. The human-capital model in Romer=s
Chapter 3 assumes that education is an investment in human capital that enhances
workers’ productivity. However, to the extent that education is mainly a signaling
tool, this model may overstate the benefits of education.

C. Understanding Romer=s Chapter 3, Part A


Introduction
Chapter 3 examines several strains of modern research literature on economic
growth. The first approach models the production of improvements in technology by
including “knowledge capital” along with physical capital. A two-sector model is
required because knowledge production does not follow the same production func-
tion as goods production; there is an R&D (or knowledge production) sector along-
side the usual sector producing physical goods.
The introduction of a second sector requires the use of some new modeling tech-
niques. For example, aggregate resources must now be divided between the produc-
tion of “goods”—either physical capital or consumption goods—in one sector and
the production of knowledge or human capital in the other. This is the role of the a
coefficients in Romer’s R&D model.
The crucial novelty of these models that makes their conclusions strongly differ-
ent than the ones of the earlier chapters is that the introduction of human or know-
ledge capital may allow us to sidestep the usual assumptions of diminishing returns
to capital and constant overall returns to scale. While it is intuitively clear that add-
ing more physical capital to a given amount of labor must eventually lead to a dimi-
nishing marginal product of capital, there is no obvious reason why increases in
knowledge would be subject to such diminishing returns. Moreover, knowledge spil-
lovers from one producer to another may allow increasing returns to scale for the
economy as a whole, even if traditional factors (labor and physical capital) produce
with constant returns to scale for any given state of technology.

Romer’s modeling strategy and the research literature


As Romer notes early in Part A, the model he presents is a simplified version of a
family of models that evolved in the growth literature in the early 1990s. A look at
the research papers he cites on pages 101 and 102 will verify for you that he has

5–9
made several simplifying assumptions. What he has done is to describe in detail a
simple version of this class of models that preserves their essential features.
For example, Romer’s model uses a Cobb-Douglas production function. This al-
lows us to evaluate marginal products explicitly and solve growth-rate equations that
would otherwise have only implicit solutions. Similarly, he relies on the assumption
of a constant saving rate in most of Chapter 3 rather than building utility maximi-
zation into the model, though this is relaxed in section 3.5. The more general models
lead to the same qualitative conclusions, so we have gained expositional simplicity
without losing the basic logic of the model. Students who are interested in the more
general approach are strongly encouraged to explore the papers cited in Chapter 3 or
to take Economics 454: Economic Growth, in which these models are examined in
more detail.

The basic setup of the R&D model


A key difference from the previous growth models that you have encountered is
that this model has two sectors. There are two kinds of capital in the models of
Chapter 3. In the R&D model, there is physical capital, which is familiar from earlier
models, and “knowledge capital.” Since there are two stocks, or state variables, we
7
have two equations of motion and must analyze convergence jointly. Romer builds
up to this gradually by first ignoring physical capital and looking at the implications
of knowledge investment as a single state variable (in Section 3.2). He then brings
physical capital back in to create the formal two-state-variable model in Section 3.3.
With two “produced” or “reproducible” factors, an additional dimension of
choice is available to the economy. Individuals can use their labor and capital re-
sources either in the sector producing physical goods or in the sector producing
knowledge. How much of the economy’s resources will be dedicated to producing
knowledge rather than goods? This is a complicated question for several reasons. As
discussed in the previous section of this chapter, there are significant conceptual dif-
ferences between these kinds of capital that may have important implications for the

7
This is a lot like what we did with c and k in the Ramsey model, but it is a little different. In
the Ramsey model, c is a “control” variable rather than a state variable because it can jump
discretely at an instant of time. (If something happens to change their situation, consumers
can raise or lower the flow of consumption spending at time t in response.) This was crucially
important in allowing the model to converge along the saddle path. If c did not jump exactly
to the value required by the saddle path, the model would have been unstable. Stock or
“state” variables such as capital and knowledge cannot jump in the same way. Their value at
any instant depends only on past investment; they change smoothly through the equations of
motion of the model. Although we can imagine discrete jumps in these variables—perhaps a
disaster that destroys capital instantly—this would imply a momentary suspension of the eq-
uation of motion that says that depreciation is proportional to the stock.

5 – 10
incentives of the private sector to invest in them. Romer avoids this issue in the con-
struction of the model by assuming that aK and aL, the shares of capital and labor de-
voted to knowledge production, are exogenous. A more satisfactory approach (that is
taken in most of the research literature) would be to endogenize these values by ex-
amining the markets for factors of production in detail and modeling the choice of
owners of factors about the industry to which they sell their resources.

The knowledge production function


There are several features of knowledge production that are worth stressing in
this model. First, knowledge does not depreciate. From Romer=s equation (3.2), it is
clear that if aK and aL are zero (so that no resources are devoted to knowledge accu-
mulation), the stock stays constant, A (t ) = 0.
The absence of depreciation of knowledge may seem counterintuitive, since old
knowledge does not seem to be worth very much in today’s world. However, we
must distinguish between the usefulness of a specific nugget of knowledge and the exis-
tence of the nugget of knowledge itself. The usefulness may decline even if we do not
have depreciation of the aggregate stock itself. The knowledge of how to produce
1980-vintage computers is only useless today because it has been superseded by even
more modern knowledge (most of which builds on the original knowledge). A rea-
sonable way to think about the absence of aggregate depreciation of knowledge in
equation (3.2) is that technical knowledge does not disappear or wear out with use
(like physical machines do). An economy that devotes no resources to the production
8
of knowledge does not slide backwards; it merely fails to progress.
A second feature of the knowledge production function is the possibility of in-
creasing or decreasing returns to scale in its production. As Romer notes on page
102, the usual “replication” argument for aggregate constant returns to scale does not
apply to the production of knowledge. He presents reasons why returns to scale
might be either decreasing or increasing.
Finally, the role of the θ parameter in equation (3.2) is very important. (This θ is
totally unrelated to the θ in the CRRA utility function that we used in the previous
chapter.) To see the intuition of the role that θ plays in the analysis, divide both sides
of (3.2) by A(t) to get

A (t )
= B[aK K (t )]β [aL L (t )]γ A(t ) θ−1. (1)
A(t )

8
There have been historical instances of knowledge being lost. Before written archives of
technological literature, knowledge of techniques could die with the individuals who knew
them. Landes (1983) describes the example of the Chinese water clock constructed by Su
Sung in 1094.

5 – 11
Equation (1) expresses the knowledge production function in terms of the growth
rate of technology (in percentage terms). Suppose that the amounts of capital and
labor allocated to knowledge production are fixed, i.e., aKK(t) and aLL(t) are constant.
Equation (1) shows that this will lead to a constant rate of technical progress if θ = 1,
since A(t)0 = 1 and thus A(t) vanishes from the right-hand side. This corresponds to
the kind of progress assumed in the Solow and Ramsey models: growth in A(t) at a
9
constant rate g.
If θ > 1, then θ − 1 > 0 and an increase in A(t) will cause the growth rate of A(t)
to increase, given fixed amounts of capital and labor devoted to knowledge produc-
tion. In other words, when θ > 1, the more knowledge we have, the faster the stock
of knowledge grows for a given amount of resources devoted to knowledge produc-
tion. More knowledge accelerates the growth rate, which of course raises the level of
knowledge even more rapidly, causing a further increase in the growth rate, and so
on. Not surprisingly, this condition turns out to be associated with explosively acce-
lerating growth in technology, productivity, and output.
If θ < 1, then θ − 1 < 0 and each increase in A(t) lowers the growth rate of A(t),
other factor inputs held constant. In this case, technology exhibits a kind of diminish-
ing returns with respect to its own production that is similar to that of capital in the
Solow model.
Capital-generated growth in the Solow model was limited by the fact that capital
faced diminishing returns in reproducing itself. Eventually, the economy settled into
a steady state in which capital could no longer grow relative to other factors. The
same thing happens to technology in the R&D model when θ < 1: Eventually, tech-
nology-induced growth is limited and, without growth in the non-produced factor
(labor), the economy becomes stationary (zero growth).

Analysis of the model without capital


Romer uses phase diagrams to search for a steady state in this model, just as we
did in the Solow model when we plotted k as a function of k and looked for the
point at which the curve intersected the horizontal axis. However, in the Solow anal-
ysis, we were looking for a “stationary value” of k, which was a ratio among va-
riables in the model (K/AL). The analysis represented in Romer=s Figure 3.1 is simi-
lar, but here we seek a stationary value for gA, the growth rate of technology, instead
of k.

9
However, note that in a Solow/Ramsey steady state the total quantities of labor and capital
are increasing, so if the a values are constant aKK(t) and aLL(t) would be increasing over time
and θ = 1 does not automatically lead to an equilibrium growth path similar to those models.

5 – 12
To find a stationary value of gA in the R&D model, we need to examine g A , the
change over time in the growth rate of A(t). That means that we are looking at the
change in a growth rate, which might be a little confusing at first.
Note that the sign of gA (and not the sign of g A ) tells us whether A(t) is growing
or shrinking. If gA is positive, then A is growing; if gA < 0 then it is shrinking. The
condition that g A > 0 means that the growth rate gA (whether positive or negative) is
getting larger as time passes. Similarly, the statement that g A < 0 means that the
growth rate of A(t) is getting smaller as time passes.
The intermediate case of g A = 0 is the case where the growth rate of A(t) is con-
stant. This situation could be a steady-state, constant-growth equilibrium. Thus, our
search for a steady state involves finding conditions under which g A = 0, then assess-
ing whether the economy would converge to such a state.
In order to find the steady state in which g A = 0, Romer first derives an expres-
sion for gA, which is just our equation (1) with β set to 0 to reflect the no-capital as-
sumption (Romer’s equation (3.7)). To get an expression for g A , he takes the deriva-
tive of (3.7) with respect to time to get (3.8). Since getting (3.8) from (3.7) is not ob-
vious upon inspection, let’s examine the intervening steps.
Looking closely at (3.7), BaLγ is constant over time, so it will not play an impor-
tant role in the time derivative. L(t) and A(t) both vary with respect to time, and their
powers are multiplied by each other in (3.7), so we will need to use the product rule
and the chain rule to differentiate with respect to time. Applying these rules directly
to (3.7) yields

g A (t ) = BaLγ ⎡⎣ γL (t ) γ−1 A(t ) θ−1 L (t ) + (θ − 1) L (t ) γ A(t ) θ−2 A (t ) ⎤⎦


L (t ) A (t )
= γBaLγ L (t ) γ A(t ) θ−1 + (θ − 1) BaLγ L (t ) γ A(t ) θ−1
L (t ) A(t )

L (t ) 
A(t )
= γg A (t ) + (θ − 1) g A (t ) ,
L (t ) A(t )

which simplifies to

g A (t ) = [ γn + (θ − 1) g A (t ) ] g A (t ). (2)

Equation (2) can be rewritten as Romer’s equation (3.9):

g A (t ) = γng A (t ) + (θ − 1)[ g A (t )]2 ,

5 – 13
which shows that g A (t) is a quadratic function of gA(t). That means that we get a pa-
rabola if we graph g A (t) as a function of gA(t). We can use basic algebra to examine
the characteristics of this parabola. Because there is no constant term in this quadrat-
ic function, g A (t) = 0 when gA(t) = 0 and it the parabola must pass through the ori-
gin. The slope of the function is its derivative with respect to gA(t),

dg A (t )
= γn + 2(θ − 1) g A (t ). (3)
dg A (t )

At the origin, gA(t) = 0, so the slope expression of equation (3) is γn > 0 and the func-
tion is sloping upward at the origin.
The sign of the coefficient on the squared term, which is θ − 1 in (2), determines
the convexity or concavity of the parabola. If θ − 1 > 0 then the slope of the function
is increasing from left to right and the parabola opens upward. Since it starts at the
origin with a positive slope, this means that it heads upward at an increasing rate as
shown in Romer’s Figure 3.2. If θ − 1 < 0, then the slope is decreasing and the para-
bola opens downward, reaching a maximum in the positive quadrant and intersect-
ing the horizontal axis as shown in Romer’s Figure 3.1. In the borderline case where
θ − 1 = 0, the function is a positively sloped straight line coming out of the origin, as
in Romer=s Figure 3.3. (The straight line is a special case of the parabola in which the
coefficient on the squared term is zero.)
Clearly, the decisive condition determining the shape of the parabola (and there-
fore the dynamic behavior of gA(t)) is whether θ is greater than, less than, or equal to
one. This provides mathematical support for our discussion in the previous section,
where the magnitude of θ was asserted to be very important.
When θ < 1, growth in technology is not “self-sustaining” due to diminishing re-
turns to knowledge. Past discoveries make future discoveries more costly in terms of
resources. Positive technological progress can only be sustained in this case if growth
in the labor force allows more and more labor resources to be devoted to research as
time passes. Note that if n ≤ 0, then the slope of the g A (t) function is zero or nega-
tive at the origin. With n ≤ 0 and θ < 1, the g A (t) function immediately turns down-
ward into the negative quadrant. In this case, the economy has steadily decelerating
technical progress approaching a steady state in which gA(t) = 0. Thus, we conclude
that in the case where θ < 1, only steady growth in the labor force will allow positive
technological progress in a steady state.
If θ > 1, the rate of technological progress may grow explosively. Each discovery
opens up multiplying new opportunities so that future discoveries become less costly
to find. Progress feeds on itself so strongly that growth in technology can accelerate

5 – 14
endlessly even with constant resources devoted to R&D. If n ≥ 0, there is no point to
the right of the origin at which the curve intersects the horizontal axis, so there is no
nonzero steady-state rate of technical progress.
It was noted in the previous section that if θ = 1, then the growth rate of technol-
ogy is neither enhanced nor retarded by the pre-existing level of technology. If the
labor force does not grow (n = 0) and θ = 1, then technological progress will occur at
a constant rate. Both terms on the right of equation (3) are zero, so g A (t) = 0 and
gA(t) remains at BaLγLγ, the level dictated by Romer=s equation (3.7). (Note that L
does not require a time index since it is constant when n = 0.) In this case, the line in
Romer=s Figure 3.3 coincides with the horizontal axis, meaning that any level of
technology growth seems to be a potential steady state—whatever the growth rate of
A, it will remain constant. The technology production function tells us that the
growth rate that the economy starts and remains at is BaLγLγ.
There are several key characteristics of the model with θ = 1 that make it interest-
ing to growth economists. First, this is a case where increasing the allocation of re-
sources to research (aL) leads to a higher steady-state growth rate. That is the sense in
which models of this kind are called “endogenous” growth models. An economy that
makes an economic choice to devote more of its resources to accumulating know-
ledge capital (perhaps through a policy of subsidizing R&D) will have a permanently
higher growth rate. By contrast, the Solow model predicts that economies that devote
more resources to capital accumulation (saving) will have a higher level of income,
but not a permanently higher growth rate. Thus, changes in the rate of capital in-
vestment have “growth effects” in endogenous-growth models but just “level effects”
in convergent models such as Solow’s.
Second, if we change our assumption about growth in the labor force to allow
n > 0, then increased labor input over time will result in everlasting acceleration of
technological progress in the θ = 1 case. A growing population means (for given aL)
more scientists, which means more discoveries and faster technological advance.
Since knowledge is assumed to be nonrival, each discovery is costlessly shared by all,
so it is the total amount of knowledge created that drives growth, not knowledge-
10
creation per capita.
Finally, as Romer notes on page 108, the crucial parameter in determining the
dynamics of the system is the magnitude of returns to scale to produced factors. By this
we mean “Does a doubling of only the produced factors leads to a doubling or more
or less than a doubling of production?” The Solow model had constant returns to
scale to all factors (labor and capital), but diminishing returns to the single produced

10
Some growth models have made the alternative assumption: that knowledge is strictly pri-
vate. In these cases, it is knowledge production per capita that matter for growth. In Chapter
6 we consider a paper by Peter Klenow (1998) that tests this assumption empirically.

5 – 15
factor (capital). Diminishing returns to produced factors assure that the sf(k) curve in
the Solow model is convex, making convergence to a steady state inevitable and rul-
ing out self-sustaining “endogenous” growth in the capital stock.
Since knowledge is the only produced factor in the R&D model of section 3.2,
the relevant condition for returns to scale in produced factors is whether θ is greater
than, less than, or equal to one. This is exactly the condition that we showed above
to have a decisive effect on the dynamic properties of the model. We shall find that
this is a quite general proposition in this class of models: growth is self-limiting, self-
sustaining, or explosive depending on whether returns to scale to produced factors
are decreasing, constant, or increasing.

The R&D model with capital


As noted above, the biggest methodological difference between the full R&D
model and previous models is the presence a second state variable K along with A.
We now explore the full, two-state-variable version of the R&D model that Romer
presents in section 3.3.
To search for a steady state, we now seek a point at which the growth rates of
both state variables are constant over time. In other words, in addition to seeking
conditions under which g A ( t ) = 0, we must also find conditions that lead to g K ( t )
= 0.
Since g A ( t ) and g K ( t ) will, in general, both depend on the current values of
both gA(t) and gK(t), we will have to use a two-dimensional phase diagram. Romer's
Figures 3.4 through 3.7 build such diagrams for two cases of the model. As we did
for the Ramsey model, we divide the space of possible values for gA and gK into re-
gions according to whether g A and g K are respectively positive or negative. To do
this, we plot the curves corresponding to the conditions g A = 0 and g K = 0. We then
use arrows to indicate the directions of horizontal and vertical motion from any
point.
For the general R&D model, it turns out that both of the relevant curves are up-
ward-sloping lines. The line corresponding to g K = 0 has a positive vertical intercept
and a slope of one; the line for g A = 0 has a negative vertical intercept and a slope of
(1 − θ)/β. The behavior of the system depends on the relative slopes of the two lines:
whether (1 − θ)/β is greater than, less than, or equal to one.
The g A = 0 line starts below the g K = 0 line, since the former has a negative in-
tercept and the latter a positive one. If (1 − θ)/β > 1, then the g A = 0 line has a stee-
per slope and will eventually intersect the g K = 0 line. Thus, if (1 − θ)/β > 1, the
model has a unique steady state with the growth rates of capital and technology set-

5 – 16
tling down to constant values. Alternatively, if (1 − θ)/β = 1, the lines are parallel
and if (1 − θ)/β > 1, the lines not only never intersect in the positive quadrant but are
getting farther apart as the economy moves away from the origin. In these cases,
there is no unique steady state.
The dynamic character of the model thus depends on the magnitude of (1 − θ)/β
relative to one. Note that (1 − θ)/β = 1 if and only if 1 − θ = β, or β + θ = 1. For a
given level of labor input (the non-produced factor), Romer’s equation (3.2) shows
that returns to scale in the production of new knowledge using the two produced fac-
tors K and A are measured by β + θ. Thus, our conclusion in the general R&D model
is parallel to our discussion above when there was no capital: with a steadily increas-
ing labor force, the model can converge to a steady state with constant growth only if
there are diminishing returns to the produced factors. With constant or increasing
returns to the produced factors, the growth rate accelerates indefinitely.

Returns to scale and endogenous growth


We have stressed several times in this chapter the importance of returns to scale
in determining the properties of the model. Specifically, we have said that the long-
run properties of growth models are determined by whether there are decreasing,
constant, or increasing returns to scale in the produced inputs.
Because this issue has had a profound impact on modern growth theory, it is
worth digressing to consider it in more detail. First of all, we need to be clear about
what we mean by a “produced input” or “produced factor.” A better term might be
“endogenous input” because we consider an input to be produced if it is created en-
dogenously within the model through the use of other factors of production. Since
pure labor is exogenous in all of the growth models we have studied, it is not consi-
dered a produced factor.
In the simple Solow model of Romer’s Chapter 1, advances in technology come
from outside—there is no way to reallocate resources to get faster technological
change—so the A term is not a produced input. However, in the R&D model of this
chapter, A is produced directly by labor and capital (and A itself). Adding more re-
sources to the R&D production function leads to more rapid accumulation of know-
ledge capital A. Thus, the evolution of A is endogenous and it is a produced input in
this model.
Returns to scale in the produced inputs are determined by what happens to out-
put if we multiply only the produced inputs by a positive constant λ. If output goes up
by less than a factor of λ, then we have decreasing returns in produced inputs. We
have constant returns to produced inputs is output goes up exactly by a factor of λ,
and increasing returns if it increases by more than that.
“Neoclassical” growth models such as the Solow and Ramsey models have de-
creasing returns to produced inputs, although their production functions usually have

5 – 17
constant returns to all inputs. Modern growth models have emphasized the case of
constant returns to produced inputs (implying increasing returns to scale in all inputs),
which leads to so-called endogenous growth. We saw in the R&D model that changes
in the economy’s choice parameters, such as the saving rate and the shares of inputs
devoted to R&D, lead to permanent changes in the steady-state growth rates in con-
stant-returns models. These parameters have “growth effects” on output in endogen-
ous growth models (i.e., they change the steady-state growth rate), but only “level
effects” in neoclassical models (where they affect the level, but not the slope, of the
steady-state growth path).
As noted in the introduction to this chapter, economists have found endogenous
growth models appealing for several reasons. First, they often lack the strong—and
arguably counterfactual—convergence implications of neoclassical models. Second,
many economists believe that such fundamental economic parameters as the saving
rate actually have growth effects rather than just level effects on real output.
The debate over neoclassical vs. endogenous growth models has spawned a vo-
luminous empirical literature. We shall examine a sample of this literature in Chap-
ter 6.

Scale effects in the R&D model


One characteristic of the R&D model that may seem unrealistic at first glance is
the presence of scale effects. Notice in Romer’s equation (3.21) that the growth rate of
knowledge depends positively on the level of the population. That means that econ-
omies with large populations should grow faster than smaller ones. This result may
seem surprising, but it is a direct result of the nonrival nature of knowledge in the
model.
Intuitively, the more people there are in the economy, the more people can work
on R&D. That will lead to the creation of more knowledge. Because knowledge is
nonrival, everyone can use this knowledge to increase productivity—as discussed
above, it is total knowledge that matters, not per-capita knowledge. The larger is the
population, the more scientists are producing knowledge (for everyone to use) and
the faster is economic growth.
International diffusion of knowledge is an important issue related to the possibili-
ty of scale effects. The relevant boundaries for the “economy” under consideration in
the R&D model are the boundaries at which new knowledge stops being used. If all
knowledge generated anywhere in the world is immediately used in production eve-
rywhere, then these scale effects occur on a global scale: growth in knowledge de-
pends on the world’s population. If some parts of the world economy operate in
(knowledge) isolation, then knowledge in these enclaves would grow at a slower rate
that is proportional to their own populations.

5 – 18
Based on this argument, it seems to be in the interest of every country to be inte-
grated into a world knowledge network where knowledge moves freely. However,
there are many reasons why knowledge might not transfer effectively on a worldwide
scale. In practice, the use of knowledge requires substantial human capital in the us-
ing country to understand and implement the advances that have occurred. Some
11
countries may lack the local population of engineers to apply new knowledge. It is
also likely that particular pieces of knowledge are more useful in some economies
than in others. For example, new hybrids of crops designed for temperate regions
may not help agricultural productivity in tropical areas. Advances in robotics may be
irrelevant to a labor-intensive economy where robots are not used because labor is
cheap relative to capital.
An interesting paper by Michael Kremer examined the plausibility of scale ef-
fects, as discussed by Romer in Section 3.7. Looking at an outrageously long time
span (and correspondingly imprecise data), Kremer (1993) does indeed find that
growth has been larger during periods and in places where population has been larg-
er. Historically isolated enclaves such as Australia and Tasmania grew more slowly
than large contiguous landmasses with large populations. Moreover, the overall
growth rate of the economy (proxied by growth in population) seems to have accele-
rated over the epochs of human history as the level of population has gotten larger.

11
This is related to the concept of “social capability” discussed by Moses Abramovitz (1986)
in a paper that we shall read in the next part of the course.

5 – 19
D. Understanding Romer’s Chapter 3, Part B
The specification of the human-capital model
The analysis of the human-capital model differs somewhat from that of the R&D
model above. Because the dynamics are complex, Romer focuses on the steady state,
looking at an equilibrium in which the amount of education per person is exogenous
and constant. This simplifies the analysis because we can focus only on one of the
two state variables: physical capital. Because K is the only state variable, the analysis
turns out to be a direct extension of the Solow model.
However, this treatment ignores some important questions that are treated more
carefully in the research literature. In particular, treating the level of education as an
exogenous variable makes exogenous one of the central decisions of the model. Ro-
mer’s simplification is analogous to the Solow model’s assumption of a constant sav-
ing rate, which makes the accumulation of physical capital exogenous rather than
responsive to economic incentives. A more complete specification of the model
would allow individuals to decide how much human capital to accumulate based on
the rate of return to education, just as agents in the Ramsey model decide on their
saving (accumulation of physical capital) based on the return to capital and their de-
sire for smooth consumption and for consumption now rather than later.
Romer begins the exposition with the production function described by equation
(3.45). Note that labor input seems to be missing from the production function here.
This apparent anomaly is resolved by equation (3.48), which expresses the amount of
human capital H(t) as the product of the number of workers L(t) and a productivity
factor G(E) that is related to the amount of education the representative worker has
received. Romer then makes the simplifying assumption that each additional year of
education adds the same proportional amount to a worker’s productivity, making
productivity an exponential function of education as shown by equation (3.50).

Analysis with the human-capital model


To begin the analysis, we must derive the intensive form of the production func-
tion (3.45). Since H(t) has taken the place of L(t) in the production function, it makes
sense to redefine y and k as

Y (t ) Y (t )
y (t ) ≡ = (4)
A(t ) H (t ) A (t )G ( E ) L (t )

and

5 – 20
K (t ) K (t )
k (t ) ≡ = . (5)
A (t ) H (t ) A (t )G ( E ) L (t )

With these definitions, we can write the intensive form of the Cobb-Douglas produc-
tion function as

y ( t ) = k (t ) α . (6)

If the level of education E is constant, as Romer assumes and as it would be in a


steady state; A and L are assumed to grow at constant rates as given by his equations
(3.47) and (3.49); thus the denominator of (4) and (5) grows at the constant rate
n + g, just as in the Solow model. It follows that the analysis of the intensive-form
model is identical to that of the Solow model from Chapter 1. We can move imme-
diately to write the unique, steady-state level of k as

1
⎛ s ⎞1−α
k* = ⎜ ⎟ , (7)
⎝n+ g +δ⎠

as shown by Romer on page 135. From (6), the corresponding steady-state level of y
is just

α
⎛ s ⎞1−α
y* = ⎜ ⎟ . (8)
⎝n+ g +δ⎠

In the steady state, y is constant at the level shown in equation (8), so its numera-
tor and denominator must be growing at the same rate. With G(E) constant, Y must
be growing at rate n + g corresponding to growth in A and L, and Y/L must grow at
rate g in the steady state. Unsurprisingly, the steady-state properties of the model are
entirely Solovian. The presence of G(E) simply scales the level of output per worker
(for given y*) by a constant amount that depends on the equilibrium amount of edu-
cation per worker.

Output per person vs. output per worker


In our analysis of the human-capital model so far, we have treated education as
though it were free. People receive an amount E without paying for it by forgoing
consumption of goods, labor effort, or investment in physical capital. Not surprising-
ly, our results suggest that more education is always better—why should we stop in-
creasing a variable that gives us higher income and costs us nothing? But of course
education is not really free. It consumes resources (mainly teacher and student time)
that could otherwise be used to produce consumable output.

5 – 21
One way of incorporating these costs of human capital would be to think of some
of our real output as being used up in education rather than being available for con-
sumption and physical-capital investment. We can think of this approach as empha-
sizing the tradeoff between the use of labor and capital resources in human-capital
accumulation (teachers and schools) vs. use in general production.
While this approach improves on the zero-cost assumption, it is somewhat un-
realistic because the greatest cost of education is the enormous number of student
hours that are diverted from production into human-capital accumulation. Even at
Reed, which is justifiably proud of its low student/faculty ratio, there are more than
ten times as many potential workers in the student body as on the faculty. Adjusting
for non-teaching staff only reduces this ratio to about three.
Romer incorporates the cost of education in his model by recognizing that time
spent as a student is time that you are consuming but not producing. Therefore, we
must distinguish carefully between output per worker (which is the Y/L value we
discussed above), and output per (adult) person, which he denotes by Y/N. The adult
population N is larger than the labor force L by the number of students.
We can tell an intuitive story about why the effect of an increase in education on
Y/N is going to be more complicated than the effect on Y/L. We assume that more
education makes workers more productive, so increased education must lead to
higher output per worker in the steady state. However, more education also means
that a smaller share of the population is working at any given time. Since students do
not produce anything, this means that the higher output of each worker is, at least
partially, offset by the smaller number of workers. Romer analyzes this tradeoff be-
ginning on page 136.
We are interested in looking at the behavior of Y/N and we know something
about the behavior of Y/L = y, so it makes sense to start by noting that

Y L Y L
= ⋅ = y.
N N L N

Thus, we need an expression for L/N, the share of the population that is working.
This can be a little tricky if people live infinitely long, as in the Ramsey model, so
Romer adopts a somewhat more realistic assumption: that everyone lives T years
with the first E years devoted to education and the last T − E years to working.
It might seem like we could then simply write down the ratio of working to total
population as (T − E)/T, since that is the share of each person’s life that he or she
works. However, we can only do this if the population is not growing (n = 0). In a
growing population, the young cohorts that are in education will be larger than the
corresponding cohorts that are working, so the share of the population working will
be somewhat smaller.

5 – 22
In order to calculate L/N for a growing population, we need to look explicitly at
cohort size. Romer denotes the flow of people born at time t by B(t). If the population
is to grow at rate n with a fixed life span, then the flow of births must grow at rate n
as well. Using our standard formula for continuous-time growth,

B (t ) = B (0)e nt . (9)

We can calculate the population at time t by adding up the sizes of all cohorts
born between t − T and t. We are working in continuous time, so this is an integral
rather than a traditional summation. Using integration to add up the flow of births
from t back to t − T gives

T
N (t ) = ∫ B (t − τ)d τ. (10)
τ=0

This is the first part of Romer’s equation (3.52). The B(t − τ) inside the integral is the
flow of births that happened τ periods before time t. The population at time t includes
those born between zero and T years before t, so integrating from τ = 0 to τ = T adds
up the cohorts that are still alive at t.
Applying equation (9) to period t − τ gives B(t − τ) = B(0)en(t − τ) = B(t)e−nτ. Romer
makes this substitution to get the second line of his equation (3.52). To get the final
line, he uses the rules of integrals to evaluate the integral expression. Because we
have not stressed the rules of integration, a more detailed explanation is appropriate
here.
Recall that integration involves “anti-differentiation,” finding the function whose
derivative equals the integrand. In this case, the integrand is B(t)e−nτ and we are inte-
grating with respect to τ. We can simplify the integral by noticing that B(t) does not
depend at all on τ, so it can be treated as a constant and brought outside the integral
sign:

∫ B (t ) e d τ = B (t ) ∫ e − nτd τ.
− nτ
(11)

Notice that I have temporarily suppressed the limits of integration in equation (11)
and treated it as an “indefinite integral.” We shall consider the limits of integration
in a moment.
Remember that the derivative of the exponential function was especially simple:

d (e ax )
= ae ax .
dx

The anti-derivative is likewise simple,

5 – 23
1
∫e dx = e ax ,
ax

or, in this case,

1
∫e
− nτ
d τ = − e − nτ . (12)
n

Equation (12) gives us the indefinite integral of the function. To calculate the de-
finite integral over the range τ = 0 to τ = T we subtract the value of the right-hand
side of (12) at τ = 0 from the value at τ = T. Thus,

⎛ 1 ⎞ ⎛ 1 − n⋅0 ⎞ 1 0 − nT 1 − e − nT
( )
T

τ=0
e − nτd τ = ⎜ − e − nT
⎝ n

⎟ ⎜ −
⎠ ⎝ n
e ⎟ =
⎠ n
e − e =
n
. (13)

Multiplying (13) by the B(t) term that we took outside the integral in equation (11)
yields the final expression that Romer arrives at in his equation (3.52).
The analysis of equation (3.53) to get the size of the labor force L(t) is exactly
analogous. The people in the labor force at time t are the members of the population
more than E years old, since people spend their first E years in education. Therefore,
people born between t − E and t are in school and people born between t − T and
t − E are working. Equation (3.53) differs from equation (3.52) only in the lower limit
of integration: τ ranges from E to T rather than from 0 to T.
Having derived the ratio of L to N in equation (3.54), Romer then proceeds to
show in a straightforward way how output per person behaves. The principal conclu-
sion was noted above. Increases in education have ambiguous effects on output per
person. They increase output per worker but decrease ratio of workers to persons.

The specification of Romer’s predation model


One of the interesting attempts to explain cross-country differences in incomes
and growth has been the development of a class of models incorporating non-
productive uses of resources. These models were inspired by the observation that
low-income economies not only usually have less physical and human capital re-
sources, but they often do not use the resources that they have productively. The ex-
perience of post-Communist Russia, in particular, has demonstrated the devastating
economic effects that can result when talented individuals find it advantageous to use
their resources in non-productive ways.
In Romer’s version of the model, individuals choose whether to apply their labor
to production or to “predation,” which involves diverting to themselves output pro-
duced by others. A more common economic term for predatory behavior is rent-
seeking. Rent seekers try to make money by various schemes of redistribution (includ-

5 – 24
ing both illegal methods like theft and such legal activities as lobbying and collecting
government transfers) rather than by producing and selling goods and services. The
presence of predation in the model leads to a third use of resources: protecting pro-
ducers from predators, including potentially both private expenditures on protection
(walls, locks, guards, and alarm systems) and social institutions such as police and
courts.
As usual, we seek the simplest possible model in which we can see the particular
effects of interest. There is no capital in the model, so it really is not a growth model
per se. Rather it aims to explain cross-country differences in output between countries
at a point in time (given their capital stocks), based on differences in predatory beha-
vior. On page 160, Romer discusses some of the reinforcing effects that enter the
model when capital is included.
One unfortunate feature of Romer’s model in section 3.11 is that he reuses f, L,
and R to represent entirely different variables than those they represented in earlier
models. In this section, f is the fraction of a producer’s resources that is spent on pro-
tection, L is the fraction of the producer’s output that is lost to rent-seeking or preda-
tion, and R is the fraction of the population that chooses to be predators.
We choose the simplest possible production function: output equals labor input.
This makes things easy by allowing us to talk interchangeably about output or labor
resources, since both are measured in equivalent units. On page 155, Romer presents
some assumptions about how producers’ losses are affected by the number of preda-
tors and the amount of resources spent on protection. These are represented by the
partial derivatives of the L function. All of these assumptions are consistent with
one’s common sense about how predation and protection should work.

Agents’ decisions in the predation model


Each person in the model must decide whether to be a producer or a predator.
This decision will depend on the returns that he or she can earn from each activity,
so we must begin by deriving an expression for the amount earned in each role.
We begin by looking at the behavior and earnings of producers. An agent who
has decided to be a producer faces a second decision: how much of his or her time to
employ protecting output rather than producing it. The more time that is spent in
protection, the larger is the share of output that is retained, but the smaller is the
overall amount produced. In other words, increasing protection activity yields the
producer a larger fraction of a smaller pie.
If there were no predation, then a producer with one unit of labor input would
get one unit of output with no need for protection. (Recall the simple one-to-one
production function we have assumed.) With predation and protection, only a frac-
tion (1 − f ) of the producer’s time is spent on production and only the share
(1 − L(f, R)) of each unit of output produced is retained after predators capture

5 – 25
L(f, R). Romer’s equation (3.62) shows how the amount of output the producer keeps
is related to f and R, the amount of his or her time spent on protection and the share
of the population that chooses predation. Producers will choose the level of f that
maximizes equation (3.62), so we take the derivative of (3.62) with respect to f and
set it equal to zero. The result is equation (3.64), which gives f as an implicit function
12
of R.
Common sense suggests that an increase in the number of predators will lead to
an increase in producers’ protection efforts. We can show that this is true by calculat-
ing df/dR, the effect of a change in R on f. If we could solve (3.64) for an explicit
function with f appearing only on the left-hand side, this derivative could be calcu-
lated in the usual way. However, with f as an implicit function, we can still find
df/dR by using the technique of implicit differentiation.
To perform implicit differentiation, we take the derivative of both sides of either
(3.63) with respect to R. We can choose to work with either (3.63) or (3.64) because
they are equivalent, but (3.63) is easier to differentiate because it does not involve
quotients. If f and R were independent, then we could take a partial derivative with
respect to R and treat f as a constant. However, because f depends on R, each time
we take the derivative of f with respect to R we get df/dR.
To see how implicit differentiation works, let’s call the left-hand side of Romer’s
equation (3.62) Z(f, R):

Z ( f , R ) ≡ −[1 − L ( f , R )] − (1 − f ) L f ( f , R ).

The rule for implicit differentiation when f is a function of R is that

dZ ( f , R ) ∂Z ( f , R ) df ∂Z ( f , R )
= + . (14)
dR ∂f dR ∂R

The first term in (14) applies the chain rule, recognizing that Z is a function of f and
that f is a function of R. Note that there are very similar looking expressions on the
left and right sides of equation (14) that are actually quite different. The total deriva-
tive on the left side is the total effect of R on Z taking all channels into account. The
second term on the right side is a partial derivative (with the ∂ symbol rather than d),
which is the effect of R on Z holding the other argument of Z (in this case, f ) con-
stant. In English, equation (14) says that the total effect of R on Z is the partial (indi-
rect) effect of f on Z times the effect of R on f, plus the partial (direct) effect of R on Z.

12
An implicit function is one where we cannot solve to isolate the dependent variable alone
on the left-hand side of the equation. Because f is “inside” the L function, whose form is un-
specified, there is no way to get f alone in equation (3.63).

5 – 26
To evaluate the effect of R on f in equation (3.63), we use the rules of differentia-
tion to evaluate the two partial derivatives in (14), using the function on the left side
of (3.63) as the Z function. This leads to Romer’s equation (3.65). Solving (3.65) for
df/dR yields equation (3.66), which is the expression we seek.
To establish that an increase in the number of predators raises protection effort,
we must determine the sign of the right side of (3.66). The assumptions that Romer
13
makes on page 155 are almost sufficient to do so. In particular, LR > 0, (1 − f ) is pos-
itive, and LfR < 0, so the numerator is a positive minus a negative, which is surely
positive. Since Lff > 0 and Lf < 0, the denominator is also a positive number minus a
negative one. With numerator and denominator both positive, the quotient is also
positive and df/dR > 0. Thus we can write the relationship between f and R as a new
function f = f(R), with f′ > 0. The amount of time devoted to protection depends posi-
tively on the number of rent-seeking predators.
Having determined the nature of the producer’s optimal protection behavior, we
next turn to the decision of whether an individual should be a producer or a predator.
As usual in economics, we assume that individuals will choose whichever alternative
14
gives them the higher return. If, given the current number of predators and produc-
ers, the return on production was higher (lower) than that on predation, then some
agents would shift from predation to production (production to predation). Thus, the
only possible (interior) equilibrium is where the returns on the two activities are
15
equal.
Romer’s (3.67) equates the returns to the two activities. The figures on pages 157
and 158 show some possible patterns of equilibria, including the interesting case of
multiple equilibria in Figure 3.11.
In order for the predation model to be useful in explaining cross-country income
differentials, we must be able to explain how differences in the amount of resources
devoted to production, (1 − f ) (1 − R), depend on institutional characteristics of an
economy. Romer does this on pages 158 and 159 by considering the effects of differ-
ences in the ease of predation. He puts this in terms of the probability of a predator
being caught and having his or her earnings taken away. This would be appropriate
when the predatory rent-seeking in question is illegal. However, we could equally

13
If we strengthen the assumptions to rule out zero partial derivatives, as we shall do, then
df/dR is surely positive. With the weak inequality assumptions a zero value is possible.
14
That means that people in this model have no “conscience”—there is no disutility attached
to being a predator.
15
A corner solution would occur if the returns to production were higher (lower) than those
to predation even if there were no predators (producers) in the economy. Then no one would
choose to be a predator (producer). There can be no corner solution in which all are predators
because there would be nothing to steal.

5 – 27
well apply this model to legal forms of predation by simply varying the ease with
which rent-seekers can claim others’ resources. For example, a very generous system
of unemployment compensation or welfare could be thought of as an easy-predation
regime, as could a political system in which advantages such as monopoly franchises
can easily be obtained by lobbying or bribes.
The discussion on pages 158 through 160 demonstrates that a decrease in the re-
turns to predation leads to a multiplier effect on output. An initial decrease in the
number of predators raises the returns to both production and predation, but (the
way Figure 3.12 is drawn) increases production returns by more, so more people will
shift from predation to production. Moreover, when fewer resources are devoted to
predation, producers will also be able to spend more time producing and less time
protecting their output from predators.
An additional, very important effect of predation is described on page 160. Indi-
viduals are extremely reluctant to invest in capital in an economy where predation
may expropriate it at some future date. This is the reason for the vast “capital flight”
from countries in which property rights are insecure. Making prospective investors,
whether foreigners or domestic residents who are currently sending their funds
abroad, feel more secure about property rights in their investments is likely to in-
crease capital accumulation, leading to higher incomes.

E. Suggestions for Further Reading


General texts on modern growth theory
Barro, Robert J., and Xavier Sala-i-Martin, Economic Growth, 2nd ed., Cambridge,
Mass.: MIT Press, 2004, Chapters 4 through 7.
Aghion, Philippe, and Peter Howitt, Endogenous Growth Theory Cambridge, Mass.:
MIT Press, 1998.
Acemoglu, Daron, Introduction to Modern Economic Growth, Princeton, N.J.: Princeton
University Press, 2009. (An encyclopedic and highly mathematical new textbook
on growth theory.)
Jones, Charles I., Introduction to Economic Growth (New York: W.W. Norton, 1998),
Chapters 4 through 8. (A much simpler mathematical treatment.)

Selected seminal papers in modern growth theory


Romer, Paul M., “Increasing Returns and Long-Run Growth,” Journal of Political
Economy 94(5), October 1986, 1002−1037. (The paper that is generally regarded
as having started it all.)

5 – 28
Lucas, Robert E., Jr., “On the Mechanics of Development Planning,” Journal of
Monetary Economics 22(1), July 1988, 3–42. (Lucas’s Nobel address focused on a
two-sector model with human capital.)
Romer, Paul M., “Endogenous Technical Change,” Journal of Political Economy 98(5),
October 1990, Part II, S71–S102. (The Paul Romer model closest to the R&D
model in David Romer’s Chapter 3.)
Aghion, Philippe, and Peter Howitt, “A Model of Growth Through Creative De-
struction,” Econometrica 60(2), March 1992, 323–351. (A seminal paper in the
neo-Schumpeterian strain of endogenous growth models.)
Grossman, Gene M., and Elhanan Helpman, Innovation and Growth in the Global
Economy (Cambridge, Mass.: MIT Press, 1991). (A book elaborating on a variety
of endogenous growth models in an international context.)
Rebelo, Sergio, “Long-Run Policy Analysis and Long-Run Growth,” Journal of Politi-
cal Economy 99(3), June 1991, 500–521. (One of the most often cited AK models.)

F. Works Cited in Text

Abramovitz, Moses. 1986. Catching Up, Forging Ahead, and Falling Behind. Journal
of Economic History 46 (2):385-406.
Acemoglu, Daron. 2009. Introduction to Modern Economic Growth. Princeton, N.J.:
Princeton University Press.
Barro, Robert J., and Xavier Sala-i-Martin. 2004. Economic Growth. 2nd ed. Cam-
bridge, Mass.: MIT Press.
Klenow, Peter J. 1998. Ideas versus Rival Human Capital: Industry Evidence on
Growth Models. Journal of Monetary Economics 42 (1):3-23.
Kremer, Michael. 1993. Population Growth and Technological Change: One Million
B.C. to 1990. Quarterly Journal of Economics 108 (3):681-716.
Landes, David S. 1983. Revolution in Time: Clocks and the Making of the Modern World.
Cambridge, Mass.: Harvard University Press.

5 – 29